A Memristor Based Deep Learning Classification Model for Object Detection

Abstract

A neural network's weight matrix is worked on via a number of training procedures. All weights will be changed simultaneously if neural networks are embedded in hardware. Neural networks that use CMOS technology, however, struggle with the weight-updating phase. Back propagation is a computationally intensive technique, but is nevertheless widely used in neural networks. However, it is much more difficult to implement at the circuit level due to the derived activation function and error back propagation. Memristor-based synaptic weight is regarded one of the essential components of parallel distributed processing, and it is believed to be on the same axis as neuromorphic computation. Despite this, however, the physical characteristics of memristors, such as material-related physics, have not yet been fully developed. Memristors are only one of several developments underway in the field of artificial intelligence (AI) in which CMOS-based Graphics Processing Unit, Field Programmable Gate Array, and Application Specific Integrated Circuit chips are being created for rapid computing. The use of memristor crossbar arrays to speed calculations is a potential method for implementing algorithms in deep neural networks in a more efficient manner. Initially, however, demonstrations have been constrained to simulations or small-scale issues, mainly owing to material and technology limitations that limit the size of memristor crossbar arrays that can be safely designed to remain stable is the objective of the work. The artificial intelligence consumes more data, noise generation from laptops, processor and memory bus’s performance will degrade while using for the training purpose, heating of the processor and other components, more CPU processing cycle are consumed and training time increases when the existing memory is employed for training the model. To address this problem the memristor can be employed as it not having any rotating motor for storing memory hence there will be no noise or heating effects. Before implementing the memristor based memory at circuit level the performance need to be simulated hence the proposed system simulates the memristor based neural network. The memristor-based neural network takes full use of the benefits of memristive devices, such as their low power consumption, high integration density, and great network recognition capacity. Designed the Neuromorphic network with nine layers of classification is used to classify binary images into 10 categories in this demonstration. Using the MNIST data set, the performance of the architecture is validated, and the impact of device yield and resistance fluctuations under various neuron configurations on network performance is also investigated. The findings indicate that the nine-layer network has an accuracy of about 98 percent in digits recognition. The simulation was carried out using, MemTorch, an open-source framework for customised large-scale memristive DL applications. The restive random access memory (Memristor based memory) is implemented using the memtorch in python3.7 scripting language. The results prove that, the same accuracy can be achieved while employing the existing memory device by considering the training time of less than a minute. Finally, a neuromorphic memristor based predicator is implemented and results show that it can be readily applied to the real time environment.